SkillRepo
Discover curated AI agent skills
Skills are quietly becoming the new standard for how people build AI workflows — packaged, reusable units of capability you snap into Claude / Cursor / Codex / Codex CLI to extend what the agent can actually do. The ecosystem around them (MCP servers, plugins, CLIs) is exploding alongside. And it's a mess.
The premise nobody seems to be solving: The discovery story is awful. Existing platforms list, they don't curate. Stars don't predict quality. And there's no way to compare two skills side-by-side without installing both.
## The interesting rabbit hole
Building the AI evaluator was where most of the time went. V1 was one prompt: "rate this skill 1–10." Everything came back 7.0. Useless. Took a few iterations to realize the LLM needs to first classify what the artifact even is (standalone skill? plugin bundle? CLI wrapper? README-only demo?) and then score it against the right rubric. Different criteria for different shapes.
The second rabbit hole: what even counts as a skill, when a plugin can contain multiple skills, a CLI can wrap an MCP server, and a lot of "awesome-list" entries are just markdown. Spent more time on the composition model than the UI.
## What's live today
Catalog with filters (type, quality score, tags)
Detail page per extension with an AI-written review
~10k entries — intentionally small while I iterate on the experience
Tech, if you care: Nuxt 4, Convex, Nuxt UI, ISR on Vercel, PostHog (no cookies)
## Where I'd love a brutal outside take
Catalog size is small on purpose — I want eyes on the experience before I scale ingestion. Even one-line reactions help more than you'd think:
Does "curated + rated" actually feel different from the awesome-lists you already ignore?
What signal would you need to trust a recommendation? Score? Reviews? Maintainer track record? A sandbox test?
On a detail page: what's missing to let you decide "install or skip" in 30 seconds?
Which sources do I have to cover beyond the official Anthropic / Cursor lists?
Honest answer — would you actually use this? If not, what's THE one feature that would flip you?
Replies like "doesn't solve problem X" or "scoring methodology is unclear" are gold. Don't sugarcoat.
Live: https://skill-repo.io
Comment